AUTOMATIC EXTRACTION OF KNEE ALIGNMENT AND MORPHOLOGY MEASURES FROM 3D MODELS IN A YOUNG-ADOLESCENT OPEN-POPULATIONS COHORT STUDY

Abstract

INTRODUCTION

Proper knee alignment is crucial for knee joint function. Little is known about knee alignment and morphology during growth; most research and current normal values were determined in adults. Imaging-based landmarks have to be identified to determine knee alignment parameters such as bisect offset or patellar translation. Currently, these landmarks are often determined manually on 2D image slices, which is time-consuming and can lead to interrater variability. Automatic extraction of these landmarks in 3D could help overcome these inconsistencies.

OBJECTIVE

To determine the concurrent validity of automatically extracted alignment parameters and morphology measures from two previously developed 3D statistical shape models (SSMs) - one for the patella and one for the distal femur- and to establish normative values and evaluate sex-based differences in these parameters among a young adolescent population.

METHODS

We included data from 1912 participants (aged 14.1 ± 0.67) who underwent knee MRI in the Generation R study, a large prospective population cohort study that follows children from fetal life until adulthood. MRI was performed using a 3.0T MRI (Discovery MR750w, GE Healthcare, Milwaukee, WI, USA), with both knees fully extended, using a water excitation Gradient Recalled Acquisition in Steady State sequence. Using a combined multi-atlas and appearance-based segmentation technique, 3638 patellae and 3355 femora were segmented from MRI scans. The 3D reconstructed bone samples derived from these segmentations were used to create two separate SSMs: one for the patella and one for the distal femur. Six patella and ten femur landmarks were annotated on the mean patella and femur shapes. Using the automatically established correspondences across bone samples during the SSMs generation, the landmarks identified on the mean bone shapes were transferred to the individual bone samples used to build the SSMs. One researcher manually annotated 30 randomly selected MRIs twice (15 boys and 15 girls) to determine the reliability of landmarks automatically extracted from the SSMs. Using these landmarks, we calculated 17 alignment parameters and morphology measurements: bisect offset; epicondylar width; femoral notch depth; femoral notch width; medial and lateral inclination angles; lateral patellar tilt; medial and lateral anterior-posterior (AP) length to epicondylar width ratio; patellar lateral translation; patellar length, thickness, and width; patellar tilt angle; sulcus angle; sulcus depth; and trochlear angle. Inter-method concurrent validity between the manually annotated parameters (mean of the two annotations) and automatically calculated parameters was determined using the intraclass correlation coefficient (ICC) for absolute agreement, calculated with a two-way mixed-effects model for single rater measurements. For alignment and morphology parameters with an ICC > 0.75, reference values (mean (SD)) and differences between these parameters in boys and girls were determined using a two-tailed t-test.

RESULTS

Six of the 17 calculated parameters, using landmarks from the SSMs, demonstrated reliable agreement, with an ICC>0.75 for bisect offset (ICC=0.86), epicondylar width (ICC=0.91), patellar width (ICC=0.75), femoral notch width (ICC=0.82), medial (ICC=0.88) and lateral condyle thickness to epicondylar width ratio (ICC=0.85). All six alignment parameters differed significantly between boys and girls (Table 1).

CONCLUSION

Only about a third of the alignment parameters and morphology measures calculated could be determined reliably. One reason might be that the position of two specific landmarks, i.e., the trochlear groove's deepest point and the patella's most posterior point, strongly influences the calculated angles. In 2D analyses, the deepest point of the trochlear groove is annotated on the same slice as the most anterior and posterior points of the femoral condyles. In 3D, these landmarks were not annotated on a 2D slice but in 3D space, resulting in different angles and distances. The low ICCs may not necessarily indicate that the 3D measurements are incorrect; they might even be more accurate; they are not directly comparable to the current clinically used 2D measurements on conventional imaging.